Annotation alignment: Comparing LLM and human annotations of conversational safety (2024.emnlp-main)
Copied to clipboard
| Challenge: | We examine whether LLMs and humans agree when annotating the safety of user-chatbot conversations. |
| Approach: | They leverage a recent DICES dataset in which 350 conversations are each rated for safety by 112 annotators spanning 10 race-gender groups. |
| Outcome: | The LLMs annotators are compared to human annotator demographic groups and can predict when one group finds a conversation unsafe . |
Similar Papers
Robustness and Confounders in the Demographic Alignment of LLMs with Human Perceptions of Offensiveness (2025.findings-acl)
Copied to clipboard
| Challenge: | despite evidence of demographic bias, reports with whom they align best are hard to generalize or contradictory . confounders introduced in the annotation process account for more variation in alignment patterns than demographic traits . |
| Approach: | They examine the alignment of large language models with human annotations in offensive language datasets. |
| Outcome: | The results show that LLMs align better with human annotations than other models. |
Aligning Language Models to User Opinions (2023.findings-emnlp)
Copied to clipboard
| Challenge: | Personality is a defining feature of human beings, shaped by a complex interplay of demographic characteristics, moral principles, and social experiences. |
| Approach: | They use public opinion surveys to model past user opinions in addition to user demographics and ideology to achieve up to 7 points accuracy gains in predicting public opinions from survey questions. |
| Outcome: | The proposed model achieves 7 points accuracy gains in predicting public opinions from public opinion surveys across a broad set of topics. |
Confident, Calibrated, or Complicit: Safety Alignment and Ideological Bias in LLM Hate Speech Detection (2026.acl-long)
Copied to clipboard
| Challenge: | censored models outperform uncensoreed counterparts in accuracy and robustness, achieving 69.0% accuracy versus 64.1% strict accuracy. |
| Approach: | They examine how large language models with minimal safety alignment compare with more heavily aligned counterparts when deployed using political personas. |
| Outcome: | The proposed model outperforms uncensored models in accuracy and robustness, while uncensors are more malleable to ideological framing. |
Evaluating Behavioral Alignment in Conflict Dialogue: A Multi-Dimensional Comparison of LLM Agents and Humans (2025.emnlp-main)
Copied to clipboard
| Challenge: | Large Language Models (LLMs) are increasingly used in socially complex, interaction-driven tasks, yet their ability to mirror human behavior in emotionally and strategically complex contexts remains underexplored. |
| Approach: | They examine alignment of personality-prompted Large Language Models in conflict dialogues that incorporate negotiation by simulating a five-factor personality profile. |
| Outcome: | The proposed model achieves the closest alignment with humans in linguistic style and emotional dynamics while Claude-3.7-Sonnet best reflects strategic behavior. |
Dissecting Human and LLM Preferences (2024.acl-long)
Copied to clipboard
| Challenge: | a recent study shows that human and Large Language Model preferences are important for model fine-tuning and evaluation. |
| Approach: | They dissect the preferences of human and 32 different Large Language Models to understand their quantitative composition. |
| Outcome: | The proposed model is compared with 32 different large language models using real-world user-model conversations. |
Fake Alignment: Are LLMs Really Aligned Well? (2024.naacl-long)
Copied to clipboard
Yixu Wang, Yan Teng, Kexin Huang, Chengqi Lyu, Songyang Zhang, Wenwei Zhang, Xingjun Ma, Yu-Gang Jiang, Yu Qiao, Yingchun Wang
| Challenge: | Existing studies on large language models have shown that they are poorly aligned in practice. |
| Approach: | They propose a framework to evaluate safety in large language models . they propose two new metrics to quantify fake alignment and obtain corrected performance estimation. |
| Outcome: | The proposed framework and two metrics show that some models with purported safety are poorly aligned in practice. |
Do Large Language Models Reflect Demographic Pluralism in Safety? (2026.findings-eacl)
Copied to clipboard
Usman Naseem, Gautam Siddharth Kashyap, Sushant Kumar Ray, Rafiq Ali, Ebad Shabbir, Abdullah Mohammad
| Challenge: | Existing datasets that focus on demographics and safety are narrow in their annotator pools. |
| Approach: | They propose to decouple value framing from responses by modeling pluralism directly at the prompt level. |
| Outcome: | Demo-SafetyBench decouples value framing from responses to model pluralism at the prompt level. |
Do LLMs Align Human Values Regarding Social Biases? Judging and Explaining Social Biases with LLMs (2025.findings-emnlp)
Copied to clipboard
| Challenge: | Large language models can lead to undesired consequences when misaligned with human values . previous studies have shown misalignment of LLMs with human value using expert-designed or agent-based emulated bias scenarios . |
| Approach: | They investigate whether large language models (LLMs) are misaligned with human values . they find no significant differences in understanding of HVSB between LLMs . |
| Outcome: | The results show that large language models do not have lower misalignment rates and attack success rates . the study also shows that smaller language models have the ability to explain HVSB . |
MEGAnno+: A Human-LLM Collaborative Annotation System (2024.eacl-demo)
Copied to clipboard
| Challenge: | Large language models (LLMs) can label data faster and cheaper than humans . however, they may fall short in understanding of complex contexts, leading to incorrect labels . |
| Approach: | They propose a collaborative approach where humans and LLMs work together to produce reliable labels. |
| Outcome: | The proposed system can produce reliable and high-quality labels faster and cheaper than humans . compared to traditional models, it can generate labels faster, at a lower cost . |
Real or Robotic? Assessing Whether LLMs Accurately Simulate Qualities of Human Responses in Human-LLM Dialogue (2026.findings-acl)
Copied to clipboard
Jonathan Ivey, Shivani Kumar, Jiayu Liu, Hua Shen, Sushrita Rakshit, Rohan Raju, Haotian Zhang, Aparna Ananthasubramaniam, Junghwan Kim, Bowen Yi, Dustin Wright, Abraham Israeli, Anders Giovanni Møller, Lechen Zhang, David Jurgens
| Challenge: | Recent work has sought to use large language models to simulate human-human and human-LLM interactions. |
| Approach: | They use a large-scale dataset to generate a paired LLM-LLM and human-LLm dialogues from the WildChat dataset and quantify how well they align with their human counterparts. |
| Outcome: | The proposed models perform similarly in simulating English, Chinese, and Russian dialogues. |